Three-dimensional residual convolutional neural networks for turbulence modeling in large-eddy simulations of partially premixed swirl flames

The partially premixed swirl flames with pressure expansion, abrupt density reduction, and stiff viscosity increases are inadequately captured by traditional eddy-viscosity models or conventional data-driven approaches such as fully connected neural networks. In this work, an advanced three-dimensional residual convolutional neural network (3D-ResCNN) is developed to accurately model subgrid-scale (SGS) stresses and scalar fluxes in such reacting flows. The 3D-ResCNN utilizes spatial inputs comprising first-order velocity and scalar gradients across local and adjacent grid points to effectively capture the SGS dynamics of turbulent combustion, eliminating the reliance on eddy-viscosity assumptions. In the a priori test, the 3D-ResCNN outperformed traditional models, including the Dynamic Smagorinsky Model and Velocity Gradient Model, achieve higher correlation coefficients for SGS variable predictions. Posteriori tests further validated its applicability and robustness within LES codes, demonstrating its effectiveness through time-averaged statistical results, velocity fluctuation analysis, and scalar probability density functions. This research highlights the advantages of integrating spatial inputs with the 3D-ResCNN architecture to enhance the accuracy, generalizability, and predictive capability of SGS turbulence models in LES for turbulent combustion.